Exponential hebbian on-line learning implemented in FPGAs

Abstract

Hebbian learning is a local learning algorithm and allows an on-line adaptation of the weights. Therefore an artificial neural network with built-in hebbian learning is capable of learning on operation. This paper presents the implementation of this algorithm in a digital Field Programmable Gate Array (FPGA). Nonlinearity is introduced by applying nonlinear low-pass-filtering to all input signals as well as using exponential shaped adaptation of weights with different time constants for rising and falling. Employing a complete serial design for the data flow, the implementation of overall 8 synapses in a single FPGA device gets possible. The neuron comprises four conventional synapses with fixed weights and four hebbian synapses for exponential on-line learning. Experiments show the improved performance of this system compared with a linear solution.